Collaborative_agents_V2

Mijin-VT — Fri, 03 Apr 2026 05:47:46 +0000

🤖 Collaborative Agents v2.1
https://github.com/Mijin-VT/Collaborative_agents_V2
Python 3.10+ License: MIT Platform LM Studio

Local AI system that enables 3 models to work together as a collaborative team.

What is this?

A system that enables 3 AI models installed on your computer to work together as a team. Each AI has a specialized role and can read files, write code, execute commands, and collaborate to solve complex tasks.

✅ 100% local — Nothing is sent to the internet
✅ No accounts or subscriptions — Everything runs on your machine
✅ Secure — Agents work in an isolated folder with action confirmation
✅ Configurable — Change models, URLs, and limits without touching code

Complete system features

🤝 Agent collaboration

Feature	Description
Automatic planning	Coordinator decomposes tasks and assigns subtasks to agents
Sequential execution	Agents work 1 at a time, reading previous results
Error recovery	If an agent fails, another automatically takes its place
Accumulated context	Each agent receives the full history of previous work
Final integration	Coordinator unifies all results into a coherent response

📂 File management

Feature	Description
Read files	Agents read workspace files with path validation
Create files	Generate new files with confirmation and preview
Modify files	Overwrites with automatic backup before each change
List workspace	Explore project file and folder structure
Content cleaning	Extracts code from markdown blocks and removes unwanted markers

⚡ Command execution

Feature	Description
Run commands	Executes system commands with 2-minute timeout
Risk analysis	Classifies each command: low 🟢, medium 🟡, high 🔴
Mandatory confirmation	User must approve each command before execution
Command editing	Allows correcting the command before approval
Pipe detection	Recursively analyzes each part of a piped command

🛡️ Integrated security

Feature	Description
Isolated workspace	Agents only access `~/agents_workspace/`
Blocked paths	Blocks `C:\Windows`, `/usr/bin`, `/etc`, etc.
Banned commands	Blocks `format`, `rm -rf /`, `shutdown`, `sudo`, etc.
Automatic backups	Backup copy before modifying any file
Session logs	Every action recorded with timestamp
Cleanup on exit	Deletes logs and backups when closing the app

🎛️ External configuration

Parameter	What it controls
`lm_studio_url`	LM Studio server URL
`workspace`	Agent working folder
`max_tokens`	Maximum tokens per response
`max_context_tokens`	Maximum context tokens
`temperature`	Default temperature
`models.*.name`	Name of each model in LM Studio
`models.*.temperature`	Per-agent temperature
`forbidden_paths`	Blocked system paths
`forbidden_commands`	Blocked dangerous commands

📊 Context management

Feature	Description
Token counting	Automatically estimates tokens (~1 token ≈ 3.5 characters)
Smart truncation	Preserves system prompt, removes oldest messages
Context recovery	If limit exceeded, reduces and retries automatically
Automatic retries	Up to 2 retries with 3-second wait

🎨 Visual interface

Element	Description
Decorative panels	Unicode-bordered boxes around text
Progress bars	Visual indicator `[████████░░░░] 66%` per phase
Animated bars	Rotating spinner + estimated progress during generation
Per-agent colors	Coordinator (cyan), Analyst (magenta), Developer (blue)
Loading indicators	"⏳ COORDINATOR thinking..."
Success indicators	"✅ COORDINATOR responded in 2.3s"
Per-flow icons	🚀 Full, 💻 Code, 💬 Debate, 💬 Free chat

💬 Workflows

Workflow	Phases	Recovery
🎯 Full task	Planning → Execution → Integration	Alternative agent if fails
💻 Development	Specifications → Implementation → Review	Analyst → Coordinator, Developer → Analyst
💬 Debate	Analysis → Technical perspective → Synthesis	Individual agent fallback
💬 Free chat	Direct conversation with visual selector	N/A

🌐 Indirect web access

Feature	Description
Download with curl/wget	Agent can use `curl` or `wget` to download content
Confirmation required	Classified as medium risk, requires approval
Content as reference	Downloaded HTML is used as source for the project

Included files

File	Description
`INSTALL.bat`	Windows Installer (run as administrator)
`install_full.ps1`	Windows Installer logic (PowerShell, executed by INSTALL.bat)
`install.sh`	Linux/macOS Installer (`chmod +x install.sh && ./install.sh`)
`RUN_AGENTS.bat`	Windows Launcher (double-click to use)
`run_agents.sh`	Linux/macOS Launcher (`chmod +x run_agents.sh && ./run_agents.sh`)
`collaborative_agents_v2.py`	Main script — The brain of the system with 3 collaborative agents
`config.json`	External config — Models, URLs, workspace, tokens, banned commands
`test_agents.py`	Unit tests — 24 tests for security, analysis, and context management
`test_cleaning.py`	Cleaning tests — Verifies code extraction and content cleaning
`test_progress_bar.py`	Progress bar tests — Verifies animated progress bar
`.gitignore`	Excludes logs, backups, workspace and caches from version control
`DOCUMENTATION_EN.md`	Complete documentation in English

What's new in v2.1

🎛️ External config (config.json): Change models, URLs, workspace and limits without touching code
🔄 Error recovery: If an agent fails, another automatically takes its place
📊 Context management: Token counting and automatic truncation to prevent overflows
🐧 Linux/macOS support: install.sh and run_agents.sh scripts
🧪 Unit tests: 24 tests covering security, command analysis and context management
🎨 Improved UI: Decorative panels, animated progress bars, loading indicators, per-agent colors
🧹 Session cleanup: Deletes logs and backups on exit to prevent file accumulation
🪟 Windows compatibility: UTF-8 forced + colorama for proper color and character rendering
📁 .gitignore: Excludes logs, backups and automatically generated files

Quick start

# 1. Install (Windows: run as admin)
INSTALL.bat

# 2. Open LM Studio → Load all 3 models → Developer → Start Server

# 3. Run the system
RUN_AGENTS.bat

Main menu

Option	Function
1. Full task	All 3 agents collaborate sequentially on your task
2. Code development	Specifications → Implementation → Code review
3. Debate/Evaluation	3 perspectives on a topic with final synthesis
4. Free chat	Direct conversation with the agent of your choice
5. View workspace	Explore project files and folders
6. Exit	Clean session and close the application

System requirements

Component	Minimum	Recommended
GPU	NVIDIA 12 GB VRAM	NVIDIA RTX 4080 Super (16 GB)
RAM	32 GB	48 GB
Disk	~50 GB free	SSD NVMe
OS	Windows 10/11, Linux, macOS	Windows 11, Ubuntu 22.04+, macOS 13+

AI Models

Agent	Model	Size	Role	Temperature
🤖 Coordinator	Qwen3 4B Instruct Heretic	~8 GB	Organizes, integrates, decides	0.3
🔍 Analyst	GPT-OSS 20B Heretic	~22 GB	Reasons, plans, analyzes	0.7
💻 Developer	Qwen3 Coder 30B A3B	~15 GB	Writes code, executes scripts	0.5

If something fails

Check the troubleshooting section in DOCUMENTATION_EN.md https://github.com/Mijin-VT/Collaborative_agents_V2/blob/master/DOCUMENTATION_EN.md#12-troubleshooting

Forem: Mijin-VT